692 PROCEEDINGS OF THE IEEE, VOL. 61, NO.

6, JUNE 1973

A Digital SignalProcessing Approach to Interpolation

RONALD W. SCHAFER AND LAWRENCER.RABINER

Abrtrac+h many digital ldgnrl procerdng systems, e.g., discussion makes it abundantly clear that linear interpolation
rocodcre, modulation
and digital waveform coding is generally not appropriate for digital signalprocessing appli-
itisn~toaltertheesmplingrateofadigi~BigaatThusitis
of considerable intemst to examine the problem of interpolation of cations. We discuss the advantages of finite duration impulse
bandlimited signals from the vierpoint of digital signal processing. response (FIR) over infinite duration impulse response (IIR)
A frequencydcunaininterpretation of the interpolationprocess, digital filters for use as interpolation filters andwe discuss the
through which it is clear that interpolation is fundamentally a linear application of recently developed design techniques for FIR
atering process, is presented. filters to the design of optimum interpolation filters. These
An examination d the relative merits of 5nite duration impulse
m n se (FIR) and Mte duration impulse response @a)digital filters arecomparedtofiltersderivedfrom classical poly-
futae as interpolation 6lters indicates thatFIR filters are generally nomial interpolation formulas to illustrate the improvement
to be preferred for interpolation. It is shown that linear interpolation that can be achieved.
snd dassicnl polynomial interpolation correspond to the use of the
FIR interpolation filter. The use of classical interpolation me&& 11. DIGITAL SAMPLING RATE ALTERATION
in signal procedug a p p l i ~ a t i ois~illustrated by a dientssion of FIR
interpolation atere derived from the Lagrange interpolation formala. A . Sampling Contirruous- Time Signals
The limitations of these 5lters leados to a consideration of optimum Consider a continuous-time signal p ( t ) with Fourier trans-
FIR 5lters for interpolation that can be designed using linear pre form
g
r- teehniqtlea ELamples are presented to illustrate the sig-
nikant improvements that are obtained using the optimum filters
I. INTRODUCTION
8(u) = J ~ 2 ( f ) ~ ~ Y .

T
HE PROCESS of interpolation is familiar to anyone
The signal e(t) is sampled to produce the sequence
who has had occasion to "read between the lines" in a
table of mathematical functions. I n digital signal pro- z(n) = 3 ( n T ) , - CQ < n < m
cessing, interpolation is required whenever it is necessary t o
change from one sampling rate to another. For example, in where T is the sampling period. The,z transform of the se-
speech processing systems, estimatesof speech parameters are quence x(n) is defined as
often computed at a low sampling ratefor low bit-rate storage
or transmission; however, for constructing a synthetic speech
signalfrom the low bit-rate representation, the speech pa-
X(2) = 2 n--m
z(n)r-".

rametersarenormallyrequired at muchhighersampling The z transform evaluated on the unit circle X(@') will be
rates [l], [2].In such cases, the sampling rate must be in- called the Fourier transform of the sequence n(n). I t is well
creased by a digital interpolation process. As another example, known that the Fourier transform of thesequence x ( % ) is
an efficient digitalrealization of a frequency-multiplexed related to the Fourier transform of .+?(t) by [5]
single-sideband system has been obtained [3] by performing
complicatedfilteringfunctions at a low sampling rate and
simplerfunctions at the highsamplingraterequired for
grouping several channels into a frequency-multiplexed for-
mat. In this process, there is a need for both increasing and If ,(t) is bandlimited, i.e., b(w)= 0 for 101 >a, and if T <r/sl,
decreasing the sampling rate. Another example where sam- then it canbe seen from (1) that
pling rate reduction is requiredis in converting a delta modu-
lation representation of a waveform to a pulse-code modula-
tion (PCM) representation [4]..
In these and other examples, is it important to thoroughly
understand the process of interpolation from the pointof view as depicted in Fig. 1 where T = z / Q .
of digital signal processing rather than from a numerical anal- Assuming that 2 ( t ) is bandlimited, the original continuous-
ysisviewpoint. Forexample,tables of mathematicalfunc- time signal can be obtained uniquely from the samples x(n)
tionsaregenerallyconstructed so that linear interpolation through the interpolation formula
producessufficiently accurateresults,andfor cases where
linear interpotation is inadequate, there existsa great variety
of higher order polynomial interpolation formulas. In signal
processing appiications there is a great temptation to try to r
get along with linear interpolation because i t is a simple tech- -(t- KT)
nique. In this paper we present a frequency-domain interpre-
T
tation of the interpolation process in which i t is clear that In many digital signalprocessing problems, we are given a
interpolation is fundamentally a linear filtering process. This sequence z ( n ) , corresponding to sampling period T,and we
must obtain from the sequence x(n) a sequence y ( n ) = &(nT');
Manuscript received December 6, 1972; revised March 7, 1973. i.e., the sequence y(n) corresponds to sampling p ( t ) at a dif-
The authors are with Bell Laboratories, Murray Hill, N.J. 07974. ferent sampling rate. If we evaluate (2) for t = nT', we obtain
SCHAFER AND RABINER: APPROACH TU INTERPOLATION 693

Fig. 1. Illaetration of the relationship between the Fourier transform of

a continuoustime signal and the Fourier transform of the sequence
obtained by sampling with period T.

a direct relationship betweeny ( n ) and x(n), but itis clear that

such an equation is impossible to evaluate because the func-
tions sin [(s/T)(t-KT)]/[(r/T)(t-RT)] are of infinite dura- Fig. 2. Samplingratereduction (T’=227. (a) Fourier transform of
tion. Rather than simply truncate these functions, it is more original sequence x ( % ) . (b) Fourier transform of sequence y ( r ) =x(2n)
showing aliasing. (c) Fourier transform obtained after sampling rate
reasonable to design finite duration interpolators. To under- reduction of a low-paw filtered version of x(%).
stand how such interpolators can be designed and to under-
stand the limitations of classical interpolators, it is useful to
consider the frequency-domain representation of the process Therefore
of changing the sampling rate. m

8. Sampling Rate Reduction-Integer Factms Y ( Z )= w(Mn)zn.

n--a
Suppose that the desired sampling period is T I - M T . If
M is an integer, this simplyimplies that the new sequence is Since w(n) is zero except a t integer multiples of M ,we obtain
m
y ( n ) = 2(nT‘) = 1Z(nMT)
Y(2) = w(n)z-*/M
= z(Mn). n--m

T h a t is, the sequence z ( n ) is “sampled” by retaining only one

out of each groupof M consecutive samples. The values of the
sequence y are samples of the original waveform li.(t); how-
ever, these samples will uniquely determine z(t) if and only if
T‘<r/Q. This is clearly j u s t a consequence of the sampling
theorem as expressed by (1).
Since we are interested in direct relationships between the 1 M-1
sequences y ( n ) and x ( n ) , i t is instructive to derive an equation ~ ( 2 =
) - x(~-~(~T/I”zZ~/M). (3)
similar to (1) that relates the Fourier transforms of the two M 1-0

sequences. The derivation of the equation is facilitated by the Ifwe evaluate Y(e) on the unit circle, withnormalization
definition of a new sequence w ( n ) which is nonzero only at appropriate for the new sampling rate, we obtain
integer multiples of M ;that is
w(n) = z(n), n = 0, fM , +2M, .
= 0, elsewhere
where the sampling period is assumed to be T for both se- There is a clear similarity between (4) and (1).
quences. A convenient representation of w(n) is An example of sampling rate reduction by a factor of 2 is
shown in Fig. 2. The Fourier transform of x ( % ) is depicted in
Fig. 2(a) for thecase when r/2Q < T < r / Qso that

where the term in brackets may be recognized as a discrete

Fourier series representation of a periodic sequence that is one
at integer multiples of M and zero otherwise. The sequence
y(n), corresponding to sampling period T ’ = M T , is
y(n) = w ( M n ) , - < n < m.
Fig. 2(b) shows Y(ehT’) for T‘=2T.Inthis case,aliasing
occurs andi t is clear that,in general, aliasing will occur in the
process of digital sampling rate reduction unless the original
sampling period satisfies
 694 PROCEEDINGS TFIE IEEE, JUNE 1973
OF

%
TI-
MQ
where Q is the Nyquist frequency of g ( t ) . If this inequality is
satisfied, however, then
1
Y(ehT') = - X ( e i . 9
M
1
= -Z(U)
MT
1 % %
= -B(o), - -< W < - .
T' T' T'
If the original sampling period does not satisfy (S), aliasing
distortion can be avoided only by passing the sequence x(n) Fig. 3. Sampling rate increase (2'" = T / 3 ) . (a) Fourier transform of e-
through an ideal low-pass digital filter with cutoff frequency quences x(%) and v(n). (b) Fourier transform of desired output of
r / P . I t can be seen that the filter must have unit gain, since interpolation process.
(4) provides the factor 1/M needed to correct the amplitude
for thenew sampling rate. This,of course, results ina sequence in V(tiUT') that are centered at w = 2 r / T and k / T must be
y(n) corresponding to a continuous time signal p(t) which is a removed by a digital low-pass filter that rejectsall frequencies
low-pass filtered version of the original signal Z ( t ) . in the range% / T< IWI
<%/TI. Furthermore, to insure that the
amplitude is correct for sampling interval T', the gain of the
C . Sampling Rate Incrcastlnteger Factors filter must be L. That is,
If the sampling rate is increased by an integer factor L;
Y ( e W ' ) = a(ei.T')V(ej.T') = a(ei.r')X(ei"T)
then the new sampling period will be TI= TIL. Since the se-
quence x(n) provides samples of the desired sequence only at 1
intervals of L samples at the new sampling rate, the remaining = - H(e*T')8(w)
samples must be filled in by interpolation. T o see how this
T
can be done using a digital filter, consider the sequence where H(ehT') is periodic with period 21/T' and
v(n) = z ( n / L ) , tz = 0, k L , f2L, . %

= 0, otherwise.
B(ehT') = L, 10 I I -
T
The s transform of u(n) is r 1
=o, -T< <T'

-
(UI
m
V ( Z )= z(n/L)z*
Thus the idealinterpolationscheme forincreasing the
samplingraterequiresthecreation of a sequence of L - 1
= 5
n=-m
z(n)z-L* = x ( ~ L ) . (6) zero-valued samples between each value of the original se-
quence, which is then filtered with an ideal low-pass filter as
in (8).
The Fourier transform of this sequence is
v(eiof")= X(ei"T'L) D. Changing by Nonintcgw Factms
In theprevious two subsectionswe have discussed methods
= X(ehT). for increasing or decreasing the sampling rate using a linear
time-invariant digitalfilter. However, because of the fact that
Thus V(ehT') is periodic with period 2 % / T = 2 r / L T t , rather
we required that the process be entirely a discrete-time pro-
than 2r/P as is the case in general for sequences associated
cess, the new sampling period T' was restricted tobe either an
with a sampling period T'. Fig. 3(a)shows V(ehT')for the
integer multiple or submultipleof the original sampling period
case T'= T / 3 . If we wish to obtain a sequence y ( n ) such that
T . This restriction can be eased somewhat by a two-step pro-
y ( n ) = &(nT') cess involving a sampling rate increasefollowed by a decrease.
Supposethedesiredsampling period is T ' = ( M / L ) T ,
then we must insure that where M and L are integers. This does not seem to be a s i g
nificant limitation in practice since any factor canbe approxi-
1 mated as closely as desired by proper choice of M and L. The
Y(e*T') = - 8(0),
T' two-step processfor interpolating to a sampling period
P = ( M / L )T is as follows:
Assuming that 1) Increase the sampling rate by a factor of L by inter-
polation as inSectionII-C.Lettheresultingsequence be
1 % % denoted yl(n) with sampling period T I = T / L .
X ( e h 9 = - *(a), - - SUI-T
T 2 ) Decrease the sampling rate bya factor of M as in Sec-
tion II-B to obtain the desired sequence y ( n ) with sampling
then it is clear from Fig. 3(a), that the images of (l/T)k(w) period TI= MT1= (iK/L)T .
SCHAFER AND RABINER: APPROACH TO INTERPOLATION 695

erations in how the filter is realized as software or hardware.

A l l of these facets of the problem are interrelated-resulting
in arbitrary tradeoffs between accuracy of interpolation and
efficiency of realization. In this section we present some ob-
servations on filter design and realization that seem to imply
that FIR filters are the proper choice for most interpolation
problems.
A . Phase Distortion
The ideal interpolator has zero phase or at most a linear
phase corresponding to an integer number of samples of delay.
I I R filters cannot have precisely linear phase [6].In contrast,
there currently exist several techniques fordesigning optimal
F I R digital filters with precisely linear phase. These filters are
optimal in thesense that the widthof transition band between
passband and stopband is minimum for given values of pass-
band and stopband ripple and specified passband and s t o p
band cutoff frequencies [7]- [ll].These filters can be designed
with arbitrarily small values for passband ripple, stopband
ripple, and transition bandwidth, at the cost of increased im-
pulse response duration. Thus with FIRfilters, the interpola-
tion error due to phase nonlinearity canbe zero and the error
due to amplitude distortion canbe made arbitrarily small. In
the case of I I R filters, althoughextremely good amplitude
characteristics can be achieved, there will always be an inter-
polation error due to phase nonlinearity.
(C 1
Fig. 4. Sampling rate change by noninteger factor (T'= )T). (a) Fourier
B. Filter Realization
transform of x ( % ) and o(n). (b) Fourier transform for increase of sam- I I R filters have recursive realizations that are very eco-
pling rate by a factor of 3. (c) Fourier transform after sampling rate nomical in terms of computational complexity. Leaving phase
reduction by factor of 2.
considerations aside, F I R filters in general require more com-
putation to achievea given accuracy of approximation to the
This process is illustrated in Fig. 4 for T ' = $ T ; Le., for a desired amplitude response than do IIR filters. However, the
netincreaseinsamplingrate.Fig.4(a) shows theFourier particularnature of theinterpolation problem makes F I R
transform of the original sequence w(n). Fig. 4(b) shows the filters computationally competitive with IIR filters.
Fourier transform of the intermediate sequence yl(n) which Consider for convenience a zero-phase F I R filter with im-
results from filtering the sequence pulse response h(n) which is nonzero in the interval
~ ( n =) ~ ( n ) , t~ = 0, +3, f6, * *

= 0, elsewhere
where N is an odd integer.' In reducing the sampling rate by
with an ideallow-pass filter having gain3 and cutoff frequency
an integer factor,i.e., TI= MT,we may need a unity gain low-
r / T . The result of reducing the sampliqg rate of ylCn) by a
pass filter to insure that no aliasing occurs in retaining only
factor of 2 is shown in Fig. 4(c).
If M <L, i.e., there is a net increase in sampling rate as every Mth sampleof the sequencex ( n ) . In thiscase computa-
in Fig. 4,no aliasing can occur, and the interpolation filter can tion is reduced because of the nature of the desired output
have a cutoff frequency a / T . However if M > L , there is a sequence. The filtered output sequence at the original sam-
possibility of aliasing. In this case aliasing can be avoided by pling rate, defined as 9 ( n ) , is
making the cutoff frequency of the interpolation filter equal
to */TI. As in Section 11-B, the resulting output sequence
y(n) will correspond to samples of a low-pass filtered version
of the original continuous time signal. where all sequences in (9) are associated with samplingperiod
T . Clearly, all values of the sequence 9 ( n ) need not be com-
111. INTERPOLATION USINGF I R DIGITALFILTERS puted since the desired output is
The previous section makes it abundantly clear that the
process of changingthesamplingraterequires a low-pass
filter. Since i t is impossible to realize the ideal low-pass filter
that is required for exact results, we must consider digital
filters that approximate this ideal behavior. As in all filtering where y(n) is associated witha sampling period TI= MT.This
problems, there are many important considerations. A basic is in contrast to a comparable I I R filter where the computa-
consideration is the choice between filters from the class of tions requiredto realize the poles of the system function
would
F I R filters and from theclass of I I R filters. Given the type of have to be camed out at the original samplingrateeven
filter to be used, the problem of approximating the ideal low-
pass filter must be solved. Finally, there are important consid- 1 See Section III-C for a comment on why N should be odd.
6% PROCEEDINGS OF THE IEBE, JUNE 1973

though M-1 out of M samples of f(n) would be discarded. for all integer valuesof r. From this equation, we see that
A further simplification results from the fact that zero-phase
F I R filters have the property h(0) = 1
k(n) = k(- n).
Thus (loa) becomes
m Constraints of this type are rather difficult to impose on an
~ ( n=
)
((X-1)

&1
k(K)[t(nM - k) + z(nM + k)] I I R filter design procedure.
Some final comments on the choice of the length of the
+ k(O)z(nM). (lob) impulse response N are in order. First,we have assumed that
N is an odd integer;however, in general N can be either even
In thecase of sampling rate increase, the interpolated out- or odd for the FIRfilter. It can be shown [6], [12] that for N
put is obtained with sampling period T'= TIL by filtering the even, a linear-phase F I R filter must have a delay of at least
sequence one-half sample. This one-half sample delayitself corresponds
u(n) = z ( n / L ) , n = 0 , fL ,f 2 L , - to interpolation between samples; thus such a filter could not
preserve the samples of the original sequence. Although there
= 0, otherwise. may be instances where the half-sample delay may not be
In this case it is convenient to write objectionable, or may even be desirable, odd values of N a p
pear to be appropriate for most applications.
n+((N-I)/Z) A second comment regarding the choice of N concerns the
y(n) = o(K)h(n - K). ( 1 1) fact that we haveasserted that by increasing N, we can
h-((X-l)/V
achieve a better approximation in the frequency domain to
Substituting for o ( k ) results in the ideal interpolating filter. In the time domain, increasing
n+((N-l) / 2 )
N implies that more values of the original sequence are in-
y(n) = x(R/L)k(n - R ) , k / L an integer volved in the computation of a given interpolated sample and
L-n-((N-l)/Z) thus we shouldexpectincreasinglybetterresults as N in-
creases. We have observed that because the sequenceu(n) has
I c
t(n/L)+(W-l)/ZL)l

L-[(nIL)-((N-l)/2L)I
x ( k ) k ( n - KL)
where [ a ] means"thelargestintegercontainedin
though (11) suggests that computation required for each out-
put sample is proportional to N,we note that only one outof
every L samples of u(n) is nonzero. Thus we see from (12) that
the actual computation requiredis proportional to NIL. Note
that in this case the symmetry of the impulse response cannot
be exploited to reduce computation.
If an IIR filter were used, relatively little saving could be If N is slightly larger than this value; the computation
achieved. In fact, ina cascade realization, almost no computa-
tional saving could be attained by considering the
the input sequence.
Changing the sampling period according to T'= ( M / L ) T ,
requires that we first increase the sampling rate bya factor L
and then reduce it by a factor M.Clearly, the savings previ-
ously discussed could be incorporated into both steps of this
process.
C . Impulse Response Constraints
The previous discussion has presented compelling reasons
for the use of linear-phase F I R filters in interpolation. T o con- multiplications by zero.
clude this section we discuss some constraints on the impulse
response that are specific to the problem of interpolation.
Recall that the output y(n) is given by (12). A reasonable
requirement on the interpolation filter is that the values
output at the original sampling times be the sameas the origi- polation methods. Our aim
nal samples.' That is, for r an integer,
y(rL) = x(rL/L) = %(I), - m < I < 00.
Equation (12) implies that
tr+(N-1)/2Ll
y(rL) = ~ ( r )= z(K)k(rL - KL) (13)
LIr-(N-1) /2Ll
* This Is ale0 a property of (2), as can be easily verihed.
( 1 2 ) mostly zero samples, the computation is proportional toNIL,
since the impulse response always spans approximately NIL
u." Al-
nonzerosamples.Suppose that we wish toalwayshave Q
samples of the original input sequence involved in the compu-
tation of eachinterpolatedsample. Using (12) i t iseasily
shown that the length of the impulse response shouldbe
N = QL, if Q and L areodd
= QL - 1 , if either Q or L areeven. (15)
of some
interpolated samples will involve Q+ 1 original samples while
zeroes in
the rest will involve only Q. If N is slightly smaller than the
value given by (15), then the interpolation will involve either
Q or Q - 1 samples. Thus to insure that Q samples of the
original sequence are always involved, N should satisfy (15).
This constraint is satisfied, as we will see, for filters derived
from classical interpolation formulas. I t is also important in
hardwareandsoftwarerealizations of interpolationfilters
where i t allows the computations tobe structured so that one
does not have to check for zero samples in order to eliminate
IV. CLASSICALPOLYNOMIAL INTERPOLATION
In this section we apply the previous discussion of the
of the interpolation process to a study of classical polynomial inter-
is to give a frequency domain inter-
pretation of these formulas that will shed some light on their
applicability in interpolation problems arising in digital signal
processing. In the course of this discussion, we shall indicate
how to derive linear phase interpolation filters from tablesof
Lagrange coefficients. We begin with a discussion of linear
interpolation.
A . Linear Intcrpobtwn
Linear interpolation involves only two consecutive sam-
ples of the original sequence r ( n ) in the computation of an
SCHAFER AND RABINER: APPROACH TO INTERPOLATION 697

interpolated sample. Specifically, the values interpolated be- v(k 1

tween two samplesx ( 0 ) and x(1) lie on a straight line connect-
ing the two originalsamples,3 with the original samples of
t
course being preserved. Thus the equation relating the output x( 3 )
y(n), having sampling period T’= TIL,to the input sequence
x ( % ) , having sampling period T , is 1
k ’
3L

(1 - -2) + (3, Fig. 5. Linear filtering interpretation of linear interpolation.

= x(0) z(1) - 0 In < L. (16)

In order to interpret linear interpolationas a linear filter-

ing process, we must derive an impulseresponse for the linear
1.0
interpolation filter. This can be done by comparing (16) to
( 1 2 ) . T o begin, i t is clear that the length of the impulse re-
sponse must be
0.8
N=2L-l
0.7
as discussed in Section 111-C. If N were larger, more than 2
samples of x ( n ) would be involved in the computationof some w
n 0.6
a La5
of the interpolated values. Likewise, if N were smaller, only E
one sample of the input would enter into the computation of 0.5
some of the interpolated values. Using this information, we
3f
0.4
can write (12) as
y(n) = z(O)h(n) + z(l)h(n - L), 0 I n < L. (17)
Comparing (16) and (17) we see that
n
h ( n ) = 1O- L
- -n, < L
L
n
h(n - L) = - OIn<L. RADIAN FREWENCY
L
Fig. 6. Comparison of Lagrange interpolation
Thus h(n) is seen to be frequency responses for L = 5.

h(n) = 1 - I n l / L , In I <L
labelled Q = 2.) We recall that the purposeof the interpolation
= 0, otherwise. (18)
filter is to remove the images of the signal spectrum that are
Clearly h(n) satisfies the requirements of (14) since h ( 0 ) = 1 centered at integer multiples of 2 r / T , while leaving the fre-
I
and h(n) = 0 for nl >L. The lengthof the impulseresponse is quencies below r / T unaltered. I t can be seen from Fig. 6,that
linear interpolation achieves significant attenuation in only a
N=2L-l very small region around each integer multiple of 2 r / T . Spe-
cifically, attenuation is 40 d B or greater in a band of width
consistent with (15) for Q = 2 . 0.35u/T centered a t 2 r / T , k / T , etc. Thus it seems reason-
Fig. 5 depicts linear interpolationas a convolution process. able to note that linear interpolation is appropriate onlyif the
The sequenceu ( k ) and the triangular envelope of the impulse original sampling rate is many times the Nyquist rate.
response h ( n - k ) are shown for the case of T ’ = T / 5 . I t is clear
that because the impulse response has duration N=5(2)-1 B . Lagrange Interpolation
= 9 , only two nonzero samples of o ( k ) are ever coincident with Clearlylinearinterpolation will not be satisfactory in
h ( n - k ) . Also i t is clear that many digital signalprocessing applications. I n classicalnu-
y(n) = z ( n / L ) , n = 0 , k L ,+ 2 L , .*. merical analysis, the inadequaciesof linear interpolation lead
to theuse of higher order polynomials; i.e., in contrast tocon-
Thesystemfunction Corresponding tolinearinterpola- necting two points bya straight line, one finds a polynomial of
tion is degree Q - 1 that passes through Q original samples. The inter-
polated values then, are samples of this polynomial. A variety
of formulas have been derived for obtaining samples of the
polynomial directly from the samples x ( % ) , however since we
are only interested in the interpolationfilter corresponding to
This system function is plotted in Fig. 6 for L = 5 . (Curve polynomial interpolation, we shall use the most convenient
form; namely, the Lagrange interpolation formula [13], [14].
The form corresponding to interpolation with equal spacing
from sampling period T to T ’ = T / Lis
698 PROCEEDINGS OF THE IEEE, JUNE 1973

hIn 1
y(n) = AtQ(n/L)z(R), Q
(20a)
even
&((+-I) /2)
01
( ( e l )/a) 0 .S 0
y(n) = ArQ(n/L)z(k), Q odd (2%) L .5
0
0
k--((Q-l)/2)
0
again
have
chosen
where
we to consider
interpolation
around 0
some arbitrary time labelled 0. In (20), the quantities A t Q ( . )
are called Lagrange
the coefficients and
are given
by the
equa- - -a
tiOnS -L 0 L 2L

(- f)k+Q/r (a)
AkQ(f)
((7)
+ !(+- R) k ) ! ( f - k) b(r)
t
Q
+p- i), Q even 0. t
't t
bl L.5

AkQ(f)=
((T)
+ !((?) k) - R)!(f - R)
-E(1+ (y)
i-0
- i), Q odd. (21b) Ib)
Fig. 7. Impulse re8pomea for three-point (Q = 3) Lagrange interpolation.
(a) Interpolationin the interval - L<n<O. (b) Interpolation in the
interval O<n<L.
Extensivetabulations of these coefficients areavailablein
tables of mathematical functions [IS]. I t is interesting to note
that intervals. Thus
1
AkQ(t) = 0, t a n integer, and # R .
Y(n) = AkQ(n/L)z(k)
= 1, t = k. b l
1

This is a result of the fact that the interpolation polynomial = z ( k ) k ( n - RL)

passes through the Q original data samples. T h a t is, t--1

for values of n
y(kL) = z(k), for k an integer.
- L < n S O
Thus the condition of (14) is satisfied for Lagrange interpola-
tion filters. OSn<L.
In general, the formulas in (20) may be evaluated for any If we compare the above two equations,we obtain
value of n such that
k(n + L ) = A-l'(n/L)
Q even
k(n) = Ao'(n/L)

In this case the same Q original samples are involved in the
computation of all the interpolated samples in the time inter-
val spanned by the original samples. If we perform the inter-
polation with a linear filter, we can choose the length of the
filter impulse response so that Q original samples are always
involvedin thecomputation: however, a given set of Q
original samples is used only to compute L-1 interpolated
samples. Thus we can interpret the Lagrange formulas in(20)
in terms of Q- 1 different impulse responses-corresponding
to the Q-1 different interpolation intervals between the
original samples. As an example, consider the case for Q = 3.
In comparing (12) and (20b), we note that the three samples
~ ( - 1 ) x~( O ) , and x(1) can enter into the interpolation in two
k(n - L ) = A l a ( n / L ) .
If these equations are evaluated for the first intervalwe obtain
the impulse response of Fig. 7(a) where L = 5 . Likewise, Fig.
7(b) shows the impulse response corresponding to Lagrange
interpolation in the second interval. Clearly neither of these
impulse responses has linear phase, since they do not satisfy
the symmetry condition h ( n ) = h ( - n ) . Indeed, it is easy to
show that whenever Q is odd, none of the impulse responses
corresponding toLagrangeinterpolationcanhavelinear
phase. However, if Q is even, one of the Q- 1 impulse re-
sponses does have linear phase.
Q As a second example suppose Q=4. Using (20a) we obtain
2
Y(n) = Ab4 ( n / L )Z ( k ) . (22)
k=-1
 phase impulseresponse derived froma Q point Lagrange inter-
polation formula is obtained from the equations

x
dl ,I :I? I
. . . . . . . . . . . . . . . . . .
h(n + L) = A-lQ(n/L)
h(n) = AoQ(n/L)
k(n - L ) = A l Q ( n / L )
s . . . . . . . . . . . . . . . . .

where 0 In <L.
I n Fig. 6 we have plotted the frequency response of the
zero-phase interpolators derived in this manner for Q= 2,4,6,8,
and L = 5 . Since the impulseresponse duration is N = QL- 1 ,
N = 9 , 19, 29, 39 for these cases. Clearly, the effect of increas-
ing Q is to improve the frequency response of the interpolator.
Whereas the linear interpolator has 40-dB attenuation in a
bandwidth of 0.35rlT centered around each integer multiple
of 2 r / T , for the 4-point and 6-point interpolators, the band-
widths areO.7r/T and O.%/T, respectively, fora t least 40-dB
attenuation. Thus at the expense of increased computation,
we can achieve a significantly better interpolation by using a
filter derived using ( 2 3 ) .
Fig. 8. Impulse responses for four-point (Q= 4) Lagrange interpolation. The interpolation filters derived from the Lagrange inter-
polation formula achieve high attenuation in a narrow band
around integer multiples of 2*/T because the zeroes of the
we can evaluate( 1 2 ) for N = QL - 1
As in the previous example system function tend to be clustered about those frequencies.
and we obtain the equation For example, in the case of linear interpolation (Q= 2), by
2 looking at (19) we see that H(e) has a double zero on the unit
~ ( n=
) x(K)h(n - KL) circle at integer multiples of 21/T. For Q=4, we have found
k-I that there are clusters of 4 zeroes not precisely on the unit
from which by comparison to ( 2 2 ) we obtain circle but close in the vicinity of w = 2 r / T , k / T ,etc. As a
result, the attenuation close to frequencies 2 r / T and 4 a / T
h(n + L) = L 1 4 ( n / L ) is very high. This is clear from Fig. 9 , where the system func-
h(n) = A o 4 ( n / L ) tion for the case Q = 6 is plotted on a log scale. However, i t is
also clear that between 2 r / T and h / T ,the response of the
k(n - L) = A14(n/L) Lagrangeinterpolation filter leavesmuchto bedesired.
h(n - 2L) = As'(n/L). Clearly, as Q gets larger, the impulseresponse gets longer and
there aremore zeroesto distribute so as to increase the attenu-
These equations can be evaluated in the3 intervals ationandbroadentheattenuationbands.This raises the
I: - L < n S O question as to how we might design F I R digital filters so as to
make the bestuse of the filter zeroes.
11: O < n < L
111: L <n 2L V. OPTIMUMF I R INTERPOLATION FILTERS
I n practical situations, signals are often sampledat a rate
to obtain the three impulse responses shown in Fig. 8 for the that is only slightly higher than twice the Nyquist frequency
condition L = 5 . The exact values for the three impulse re- in orderto minimize thecomputationrequired for digital
sponse coefficients are also given in Fig. 8. We note that the filtering and other signal processing procedures. In this case,
impulse response corresponding to interpolationin the central the ideal interpolation filter for increasing the sampling rate
interval 0 In <L,is symmetric and thus has zerophase. The has constant gain in the frequency range 0 I W I < r / T , and I
two other impulse responses clearly do not have linear phase. zero gain elsewhere. For such signalswe are clearly interested
Indeed, it is clear that,in general, Lagrange interpolation has inthebest possible approximationstotheideal low-pass
phase distortion except in the Q even case and interpolation filter.
in the central interval. Thus if we wish to use impulse re- On the other hand, in situations where the original se-
sponses derived from the Lagrange interpolation formula for quence x ( n ) is derivedbysampling at a rate considerably
interpolation in signal processing applications we should use higher than twice the Nyquist frequency, we require a rela-
theconditionsthat yield zerophase. I n general thelinear tively narrow passbandof constant gain anda number of stop-
700 PROCEEDINGS THE IEEE, JUNE 1973
OF

must also be close to r / T . Thus the transition band wp<w

<wn, must be very narrow, which implies that a large valueof
N will be required.
-20
O m I n such cases, it is reasonable todefine only one stopband,
w,, <w < r / T ‘ . However, in cases where r / T is significantly
higher than the Nyquist frequency, the transition band be-
tween passband and stopband can be wider,and itmakes sense
to define stopbands around each integer multiple of 2 r / T ,
with transition bands w,,+hW, <w < u , ~etc.,
, in which the fre-
quency response is unconstrained. As can be seen fromFigs. 6
and 9, this is the type of frequency response that characterizes
the Lagrange interpolation filters.

B . Design Techniques
The frequency response of a zero-phase F I R filter can be
expressed as
(N-1)/ 2
E(ehT’)= h(0) +2 n-1
h(n) cos (wnT’).
(24)

RADIAN FREOUENCY The tolerance scheme depicted in Fig. 10 can be expressed as

the following set of inequalities:
Fig. 9. Frequency respome of six-point (Q = 6) linear phase Lagrange
interpolation filter. L = 5, N = 29.
1-KK621H(ehTf)51+K62, Olwlw, (25)
-6 2 IH(ehT’)I&, w~IwIw~+AwI
w . , ~ w I w , ~ + A w ~ (26)

where we have defined K = &/Zit as the ratioof the passband to

stopband error. For a given N and K ,these equations may be
evaluated on a dense set of frequencies in the specified pass-
band and stopbands and may be solved for { h ( n )] and 6,
using either linear programming techniques [ll] or discrete
Chebyshev approximation techniques[ l o ] .I t has been shown
[ l o ] that thefilters designed by these techniques are optimum
in the senseof having the narrowest transition bands for given
61 and as. For these filters, the error in the passband and stop-
bands exhibits a n equiripple behavior..An example of a low-
pass fikerdesigned by linear programmingis shown in Fig.11.
In this case L = 5 , N = 2 9 ( Q = 6 ) , w , = O . h / T , w , , = l . k / T ,
bands of zero gain, with the frequency response being some- K = 1.0. The resulting filter has 61=6r=0.00586. (The filter
what arbitrary elsewhere. This means that high-order poly- gain has been normalized to 0 dB for convenience in plotting.)
nomial interpolationfiltersmaybequitesatisfactoryfor The equiripple behavior of the stopband is readily apparent.
signals that are sufficiently oversampled. However, it is gen- The filter in this example does not satisfy the constraints
erally possible toachieve significantly betterinterpolation on h(n) given in (14); however, these constraints are linear
filters using optimization techniques. and easily can be added to the constraints of ( 2 5 ) and (26).
Thus in the optimization procedureh(0) is constrained to 1.0
A . Design Spccificotwns and h( fL ) , h( f2L), etc., are set to0. The filter performance
A scheme for approximatingthe ideal interpolatoris is only slightly degraded by the addition of these constraints.
shown in Fig. 10. There is a band of frequencies, 0 <w <wp, Fig. 1 1 shows an example where all the fixed-design parameters
called the passband inwhich the frequencyresponse should be were the same as in the previous example. In this case the
close to 1.‘ In practice we allow for an error of f61. In addi- resulting value of &=& was 0.00599. I t can be seen from Fig.
tion, there are one ormore stopbands in which the frequency 12 that the equiripple nature of the frequency response is
response is required to be within f a n of zero. The choice of destroyed, but with little sacrifice in performance.
the parameters in Fig. 10 depends upon the desired accuracy If the original signal bandwidth is much less than r / T ,
of interpolation and to some extent upon the nature a bandstop filter may provide superior performance toa com-
of the
input sequence. The choice of passband and stopbands de- parable low-passfilter. Fig. 13 shows an example for L = 5
pends upon the Nyquist frequency where N = 2 9 , w , = O . h / T , ~ . ~ = l . i r / T , A w l = & = 0 . 6 r / T ,
of the original signal. If the
original sampling period is such that r / T d l , then w p must w,,~=3.7r/T,andK=l.O.Inthiscase6~=6~=0.0000681.Thus
be very close to r / T , and the first stopband, beginning at w,,, thls filter does a much better job of attenuating the imagesof
the signal spectrum around the frequencies 2 r / T and k / T ,
but only if the original signal bandwidth was much less than
4 If the sampling rate is increased by a factor of L,the 6ltergain must
approximate L in the passband instead of 1.0. This is easilyaccomplished r / T . I t can be seen that the extra attenuation around 2 r / T
by mnltiplying the impulse response samplea by L. and k / T is obtained at the expense of the regions around
 SCHAFER AND RABINER: APPROACH TO INTERPOLATION
701

0' -
- --
LMRANGE
OPTIMUM
BANDSTOP
-400
I
- 1201 I I I I I I I I I
-
0.w
T
-
o . 2 ~-
T
ab-
T
a4r -
05*
T T
- - - -
o . 6 ~ 0 . 7 ~ OBT
T T T T
agu

9
I I I I
Fig. 14. Comparison of linear-phaseLagrange interpolators and opti-
T
0 11
T
-
3r
T
Az
T
a*+
T T
mum bandstop filters ( L= 5). Plots show minimum stopband attenua-
RADIAN FREPUENCY tion as a function of half the stopband width. For the optimum
bandstop filter (N= 29), the curve is off scale for oP<0.2r/T.
Fig. 11. Frequency response of a typical low-pass interpolation filter
designed by linear programming. L = 5, Q = 6, N = 29, up=0 . 6 r / T ,
, = 6z=O.W586. (No impulse response constraints.)
u . ~ =1 . 4 ~ / T61 31/T and 5s/T where the attenuation is much less. Thus
optimum filters of this type (and classical filters) should only
be used if one is certain of the bandwidth of the input se-
quence.

C. Comparison to Classical Interpolators

I t is of considerable interest to compare thefilters derived
from classical interpolation formulas and those designedby
I linear programming. In the design of the optimum filters, the
parameters of the tolerance schemeof Fig. 10 were set so that
& =6 2 with symmetrical stopbands of width Awl=A%= . .
= h,centered a t integer multiplesof 27r/T. This is reasonable
sincein interpolation, preservation of the passband is gen-
erally as important as rejection of stopbands. In order to com-
pare the Lagrange filters to the optimalfilters, values of 61 and
6 2 were measured at the edges of the passband and the first
stopband,respectively.Since for theLagrange filters, 61 is
I I I
T T
0 - I
T
TT
-
2a
T
-
3* 5E
I
%zsz
T'
always greater than B2 for the above definition of passband
and stopbands, the comparison hasbeen made on the basis of
RADIAN FREOUENCY passband error.
Fig. 12. Frequency responsefor the example of Fig. 1 1 , with impulse Fig. 14 shows such a comparison for the case L = 5 . 5 The
response constraints. In this case 61 = 62 = 0.00599. solid curves are for linear-phase filters derived from the La-
grange interpolation formula where N = QL - 1 and Q = 2, 4,
and 6 . These curves show the error in decibels at the edge of
the passband, i.e., 20 loglo &,as a function of up.As an ex-
ample, for linear interpolation ( Q = 2 and N = 9 ) we see that
the passband error is - 42 dB for up= O.ln/T. Furthermore,
since &>6? we can say that the attenuation is a t least 42 dB
in the region (2r/T-O.l7r/T)<u<(27r/T+O.la/T).For
wider bandwidths,theperformance is worse:however, for
higher order interpolators the performance becomes appreci-
ably better as is expected.
The dotted curvesin Fig. 14 show passband error for band-
stop filtersdesigned bylinearprogramming. By comparing
corresponding curves, it can be seen that the optimumdesigns
are always significantly better than the corresponding classi-
cal interpolator, with the improvementbeing most strikingfor
narrow bandwidths and for the higher order filters.
Clearly, there are a variety of optimum designs correspond-
- 12010 I I I I I ing to situations in which passband and stopband approxima-
T T
-
7r
T
-27r -
37r 42
T
2ISZ
T T'
tion errors are not treated as being of equal importance. For
RADIAN FREOUENCY example if & = K&, then the case K > 1 corresponds to placing
Fig. 13. Frequency response of a typical bandstop interpolation filter
more importance on stopband attenuation than on passband
designed by linearprogramming. L = 5 , Q = 6 , N = 2 9 , w , = 0 . 3 ~ / T ,
o , , = 1 . 7 r / T . o s = 3 . 7 * / T . A W = A W Z = O . ~ T /and
T , 61=6~=0.oooO681. 5 The comparisons are similarfor larger values of L.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Cw, T I * )
Fig. 15. Comparison of optimum bandstop andoptimum low-pass inter-
polationfilters(L=5).ThecurveforN=29isoffscaleforwp<0.2r/T.
error. This situation would be a more favorable situation for
the classical filters, although i t is always possible to design a
better filter using the optimum design procedures.
Another interesting comparison is between optimum low-
pass filters and optimum bandstop filters. T o compare these
filters, we set
2* 4*
* p = - - WI1 = - - *I2
T T
and & =&2= 2wp for the bandstop filters and
2*
UP=-- *a1
T ZEEE Trans. Audio Elcdrocrccncst., vol. AU-21, pp. 165-174, June
and &=T/T’--W,, for the low-pass filters. T h a t is, both filters
were designed to accommodate the same input signal band-
width. Fig. 15 shows the difference between stopband attenua-
tions for the bandstop filters and thelow-pass filters as a func-
tion of bandwidth. From these curves we see that for narrow
bandwidthsthebandstop filters have significantly greater
attenuation; however as the bandwidth approaches half the
original sampling frequency, there is no difference between the
two types of filters.
VI. CONCLUSIONS
I n this paperwe have discussed the process of interpolation
as a problem in digital filtering. Most of our discussion has
involved frequency-domain representations of the interpola-
tion process and design criteria for digital interpolation filters.
We have taken this approach because i t is the most reasonable
for digital signal processing applications where i t is necessary
to either raise or lower the sampling rate of a signal. This
point of view is in contrast to that of interpolation in tables
where one isconcerned primarily with minimizing the error in
a particular interpolated sample. Because of the variety of
factors involved in the design of an interpolation filter, we
have not tried to give design formulas and error bounds that
would have limited value, but rather we have chosen to at-
tempt to illuminate the important factors involved in the
PROCEEDINGS OF THE IEEE, JUNE 1973
interpolation process and to discuss general design procedures
that can be adapted to a variety of situations.
In particular, we have argued that linear-phase F I R filters
have many attractive features for discrete-time interpolation
and have shownhow they may be efficiently utilized. Classical
polynomial interpolation has been discussed in the context of
digital signalprocessing. Interpolation filters derivedfrom
polynomial interpolation formulas are attractive because the
impulse response can be easily computed or looked up in a
table. However, we have seen that the frequency response of
such systems leaves much to be desired in digital signal .pro-
cessing applications where the original sampling rate may be
only slightly above the Nyquist rate.
As an alternative tofilters based on classical interpolation
0% formulas, we discussed optimum low-pass and bandstop FIR
filters that were designed by linear programming. The band-
stop filters have frequency responses that are very similar to
the classical interpolators, but are always superior. The band-
stop designs appear to be most important for cases when the
original sampling rate is several times the Nyquist rate,while
the low-pass designs are appropriate when the original sam-
pling rate is close to the Nyquist rate.
ACKNOWLEDGMENT
The authors would like to acknowledge the helpful com-
mentsandcriticismsprovidedbyDr. 0. Herrmannand
Dr. J. F. Kaiser.
REFERENCES
[ l ] A. E. Roeenberg, R. W. Schater, and L. R. Rabiner, “Effects of
smoothing and quantizing the parameters of formant-coded voiced
speech,” J . Acoust. Soc. Amer., vol. 50, pp. 1532-1538, Dec. 1971.
[2] R. W. Schafer and L. R Rabiner, “Design and simulation of a speech
analysis-synthesis system based on short-time Fourieranalysis,”
1973.
[3] S.L.Freeney, R. B. Kieburtz, K. V. Mina, and S. K. Tewksbury,
“Design of digital filters for a n all digital frequency division multi-
plex-time division multiplex translator,” ZEEE Trans. Circuit Thc-
my, V O ~ .CT-18, pp. 702-711, NOV.1971.
[4] D. J. Goodman, ‘The application of delta modulation to analog-to-
PCM encoding,” Bell Syst. Tech. J . , vol. 48, pp. 321-344, Feb. 1969.
[SI H. Freeman, Discrete-Time Systems. New York: Wiley, 1965.
[a] A. J. Gibbs, ‘On the frequency-domain r e s p o m of causal digital
filters," Ph.D. dissertation, University of Wisconsin, Madison, 1969.
[7] L. R. Rabiner, ‘Techniques for designing finite-duration impulse-
response digital filteq”IEEE Trans. Commun. Tech&., vol. COM-
19, pp. 188-195, Apr. 1971.
[8] 0.Herrmann, ‘On the design of nomecursive digital filters with
linear phase,” Electron. Lett., vol. 6, pp. 328-329, 1970.
[9] E.Hofstetter, A. V. Oppenheim, and J. Siegel, “A new technique for
the design of nonrecursive digital filters,” in Proc. 5th Ann. Princdm
C a j . on Znjormation Sciences and Syskms, pp. 64-72, 1971.
[lo] T. W. Parks and J. H. McClellan, “Chebyshev approximation for
nonrecursive digital filters with linear phase,” ZEEE Trans. Circait
T h y , vol. CT-19, pp. 189-194, Mar. 1972,
[ l l ] L.R. Rabiner, “The design of finite impulee response digital filters
using linear programming techniques,” Bdl Syst. Tech. J . , vol. 51,
p ~ 1177-1198,
. July-Aug. 1972.
[12] L. R. Rabiner and R. W. Schafer, “Recursive and nonrecursive re-
ahations of digital filters designed by frequency sampling tech-
niques,” ZEEE Trans. Audio Electroocoust., vol. AU-19, pp. 200-207,
Sept. 1971.
[13] R. W. Hamming, Numerical Methods jor Scientists and Engineers.
New York: McGraw-Hill, 1962.
[14] F. B. Hildebrand, Introdaction to Nvnrcrical A d y s i s . New York:
McGraw-Hill, 1956.
[15] Handbook o j Mdhematicd Functions, M.Abramowitz and I. A.
Stegun, Eds. (Nat. Bur. Stand. Appl. Math. Ser.45.) Washington,
D. C.: U. S. Gov. printingOftice. 1964.